SVIE: Stitching Video Priors for Physics-Aware Image Editing
Abstract
Recent advancements in text-guided image editing have demonstrated remarkable capabilities in synthesizing realistic textures and following complex instructions. Nevertheless, most existing models are trained primarily on static images and inherently lack the ability to reason about physical causality. This limitation often leads to ''physical hallucinations'', where generated content violates physical laws (e.g., inconsistent lighting, implausible gravity, or multi-step causal effects). In contrast, video generation models implicitly encode rich physical priors, but they are ill-suited for high-fidelity image editing due to task misalignment and limited spatial resolution. To address this gap, we present SVIE, a novel framework that seamlessly transplants the physical reasoning capabilities of video generation models into image editing models. Technically, SVIE achieves this by stitching a frozen video generation backbone into the image editing pipeline via a specialized Stitching Adapter. This lightweight adapter acts as a semantic bridge, projecting causality-rich video latents directly into the image model's feature space. These aligned priors are then injected into the denoising process to encourage more physically consistent outputs. Extensive experiments show that SVIE substantially improves physical plausibility in challenging scenarios while preserving instruction adherence and detail fidelity. Notably, SVIE enhances physical reasoning accuracy, achieving an absolute gain of 5.2 percentage points in overall accuracy on RISEBench and consistent gains on PICABench. Crucially, it secures these gains while maintaining competitive performance on general editing benchmarks like ImgEdit.